原文 · 未翻译
Anthropic Tuesday publicly released Claude Fable 5, its first “Mythos-class” model that it says surpasses its previous frontier Opus models in overall capabilities. But the model’s launch today comes with safeguards designed to prevent it from answering queries on topics like cybersecurity, biology, and chemistry, where the company has publicly worried about its potential impact to “uplift” malicious actors.
Anthropic says Fable 5 operates on the “same underlying model” as Mythos 5, which is coming out of its monthslong “Mythos Preview” period today, but only for “a small group of cyberdefenders” judged trustworthy through the existing Project Glasswing. Unlike Mythos 5, though, the publicly accessible Fable 5 is designed to funnel queries on certain sensitive topics to the earlier Claude Opus 4.8 model and to warn the user when this is happening.
Anthropic said it has tuned these safeguards to be “stricter than ideal,” meaning the system may occasionally refuse “harmless requests” in a way that it acknowledges may be frustrating for regular users. But Anthropic says such false positives come up in less than five percent of all sessions in testing, and were worth it to avoid situations where Mythos could give malicious actors assistance in “causing serious harm that they couldn’t have received from other sources.”
I can’t let you do that, Dave
Fable 5’s topic-based safeguards are built around a system of classifiers designed to broadly detect banned prompt subjects as well as any potential jailbreak attempts. In over 1,000 hours of red-team testing with a bug bounty program, Anthropic says external teams failed to find any universal jailbreaks for Fable 5. The new model also resisted automated jailbreak attempts to a much larger degree than previous Claude Opus models, Anthropic said.
The company said it is particularly worried about Mythos 5’s ability to perform “agentic hacking,” executing multi-part cyberattacks with much more facility than earlier models. But testing from the UK’s AI Security Institute in recent months found that Mythos Preview performed similarly to OpenAI’s GPT-5.5 on a suite of Capture the Flag challenges, suggesting Mythos’ performance is not “a breakthrough specific to one model.”
Among the usual raft of fair-to-middling benchmark test improvements that Anthropic reports for Mythos 5 over previous frontier models, the company claims a significant jump in the model’s capabilities on the cybersecurity-focused ExploitBench test. Mythos 5 scored a 78 percent on the benchmark’s tests of vulnerable code exploits, a significant increase from the 40 percent score from Opus 4.8, and even the 69 percent score achieved by Mythos Preview.
While earlier Anthropic models blocked bioweapons-related queries, that classifier now applies to all chemistry and biology-related queries in Fable 5. The company says it worries that “well-resourced malicious actors” could use even seemingly benign queries on these subjects to assist with “highly risky biological research” in a much more effective way than with previous models.